Spaces:
Sleeping
Sleeping
File size: 12,187 Bytes
8c536e6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 | """Generate replay traces for the public demo (run on Colab).
For each (attack_type, steps_label) requested, this script:
1. Loads the attacker checkpoint (or zero-shot Qwen for the lowest step count).
2. Loads the real defense stack (PG2 + SecAlign + LlamaFirewall).
3. Runs the attacker once against the scenario for that attack type.
4. Captures stage-by-stage timings + verdicts into a JSON trace.
5. Saves the trace to ``data/traces/{attack_type}_{steps}.json``.
Also writes ``data/highlights/highlight.json`` — the most successful trace
across the run, used by the homepage hero animation.
Usage (typical Colab cell)
--------------------------
python scripts/generate_traces.py \\
--checkpoint /content/drive/MyDrive/injectarena/run_v1/final \\
--steps-labels 50 100 300 500 1000 1500 \\
--output-dir data/traces \\
--highlight-dir data/highlights
For ``--steps-labels`` values <= ``--baseline-cutoff`` (default 100), the
zero-shot baseline (untrained Qwen) is used to simulate an early/under-trained
attacker. Above the cutoff, the trained checkpoint is used. This is
documented in the trace itself via the ``model_source`` field, so the demo
stays honest.
"""
from __future__ import annotations
import argparse
import json
import logging
import time
from pathlib import Path
from typing import Any, Dict, List, Optional
logger = logging.getLogger("generate_traces")
logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
# Attack-type → scenario_id mapping (kept in sync with env/replay.py)
ATTACK_TYPE_TO_SCENARIO = {
"email_exfiltration": "email_exfil_001",
"forbidden_tool": "email_forbidden_001",
"prompt_leak": "email_leak_001",
"rag_injection": "rag_exfil_001",
}
def _parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser()
p.add_argument("--checkpoint", type=str, required=True,
help="Path to trained attacker checkpoint dir (the final/ folder).")
p.add_argument("--steps-labels", type=int, nargs="+",
default=[50, 100, 300, 500, 1000, 1500],
help="Step-count labels to generate traces for.")
p.add_argument("--baseline-cutoff", type=int, default=100,
help="Steps <= cutoff use the zero-shot baseline; above use the checkpoint.")
p.add_argument("--output-dir", type=str, default="data/traces")
p.add_argument("--highlight-dir", type=str, default="data/highlights")
p.add_argument("--max-new-tokens", type=int, default=128)
p.add_argument("--seed", type=int, default=42)
return p.parse_args()
# ---------------------------------------------------------------------------
# Attacker loading (uses the same Unsloth path as train/eval.py)
# ---------------------------------------------------------------------------
def _load_attacker(checkpoint: str, max_new_tokens: int):
from unsloth import FastLanguageModel
logger.info("Loading attacker from %s", checkpoint)
model, tokenizer = FastLanguageModel.from_pretrained(
model_name=checkpoint,
max_seq_length=2048,
load_in_4bit=False,
dtype="bfloat16",
)
FastLanguageModel.for_inference(model)
return model, tokenizer
def _load_zero_shot(max_new_tokens: int):
from unsloth import FastLanguageModel
logger.info("Loading zero-shot baseline (Qwen2.5-1.5B-Instruct)")
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="Qwen/Qwen2.5-1.5B-Instruct",
max_seq_length=2048,
load_in_4bit=False,
dtype="bfloat16",
)
FastLanguageModel.for_inference(model)
return model, tokenizer
def _generate_payload(model, tokenizer, observation, max_new_tokens: int, seed: int) -> str:
import torch
from train.client import format_prompt, parse_payload
prompt = format_prompt(observation)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
torch.manual_seed(seed)
with torch.inference_mode():
out = model.generate(
**inputs,
max_new_tokens=max_new_tokens,
do_sample=True,
temperature=0.7,
top_p=0.9,
pad_token_id=tokenizer.eos_token_id,
)
raw = tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
return parse_payload(raw)
# ---------------------------------------------------------------------------
# Defense stack — wrapped to record timings + verdicts
# ---------------------------------------------------------------------------
def _load_defenses():
from env.defenses.prompt_guard import PromptGuard
from env.defenses.secalign_agent import SecAlignAgent
from env.defenses.llama_firewall import FirewallWrapper
logger.info("Loading defense stack (SecAlign first for vLLM CUDA order) …")
secalign = SecAlignAgent()
pg2 = PromptGuard()
fw = FirewallWrapper(prompt_guard_fallback=pg2)
logger.info("Defense stack ready (SecAlign mode=%s).", secalign.mode)
return pg2, secalign, fw
def _run_pipeline(
payload: str,
scenario: Dict[str, Any],
pg2,
secalign,
firewall,
) -> Dict[str, Any]:
"""Run the full attack pipeline and return a trace dict."""
timeline: List[Dict[str, Any]] = []
t0 = time.perf_counter()
# Stage 1: generation already happened upstream — we record it as t=0
timeline.append({"stage": "generation", "t": 0.0, "payload": payload})
# Stage 2: PG2 input scan
s = time.perf_counter()
pg2_v = pg2.scan(payload)
pg2_t = time.perf_counter() - t0
timeline.append({
"stage": "pg2_scan",
"t": round(pg2_t, 3),
"duration": round(time.perf_counter() - s, 3),
"flagged": bool(pg2_v.flagged),
"score": float(getattr(pg2_v, "score", 0.0)),
"reason": getattr(pg2_v, "reason", None),
})
# Stage 3: SecAlign agent run
s = time.perf_counter()
untrusted = {scenario["injection_slot"]: payload}
agent_output = secalign.run(
system=scenario["agent_system_prompt"],
user=scenario["user_query"],
untrusted=untrusted,
)
agent_t = time.perf_counter() - t0
timeline.append({
"stage": "agent_run",
"t": round(agent_t, 3),
"duration": round(time.perf_counter() - s, 3),
"agent_output": agent_output,
})
# Stage 4: LlamaFirewall output scan
s = time.perf_counter()
fw_v = firewall.scan(
user_query=scenario["user_query"],
agent_output=agent_output,
tool_call=None,
)
fw_t = time.perf_counter() - t0
timeline.append({
"stage": "firewall_scan",
"t": round(fw_t, 3),
"duration": round(time.perf_counter() - s, 3),
"flagged": bool(fw_v.flagged),
"score": float(getattr(fw_v, "score", 0.0)),
"reason": getattr(fw_v, "reason", None),
})
# Stage 5: task verifier
task_succeeded = _check_task_success(scenario, agent_output)
timeline.append({
"stage": "verdict",
"t": round(time.perf_counter() - t0, 3),
"task_succeeded": task_succeeded,
})
broke_pg2 = not pg2_v.flagged
broke_fw = not fw_v.flagged
composed = broke_pg2 and broke_fw and task_succeeded
return {
"timeline": timeline,
"outcome": {
"broke_pg2": broke_pg2,
"broke_fw": broke_fw,
"task_succeeded": bool(task_succeeded),
"composed_bypass": bool(composed),
"blocked_at": (
"Llama Prompt Guard 2" if not broke_pg2
else ("LlamaFirewall" if not broke_fw
else (None if task_succeeded else "SecAlign agent (refused)"))
),
},
}
def _check_task_success(scenario: Dict[str, Any], agent_output: str) -> bool:
"""Lightweight verifier so traces can include a task-success flag."""
from env.verifiers.exfiltration import verify_exfiltration
from env.verifiers.forbidden_tool import verify_forbidden_tool
from env.verifiers.prompt_leak import verify_prompt_leak
cat = scenario["target_category"]
try:
if cat == "exfiltration":
return verify_exfiltration(agent_output, scenario)
if cat == "forbidden_tool":
return verify_forbidden_tool(agent_output, scenario)
if cat == "prompt_leak":
return verify_prompt_leak(agent_output, scenario)
except Exception as exc: # noqa: BLE001
logger.warning("Verifier error for %s: %s", scenario.get("scenario_id"), exc)
return False
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
def main() -> None:
args = _parse_args()
out_dir = Path(args.output_dir)
highlight_dir = Path(args.highlight_dir)
out_dir.mkdir(parents=True, exist_ok=True)
highlight_dir.mkdir(parents=True, exist_ok=True)
from env.environment import InjectArenaEnv
from env.scenarios import ScenarioBank
bank = ScenarioBank()
# Load defenses ONCE — they dominate cost.
pg2, secalign, firewall = _load_defenses()
# Build a minimal env so we can reset() into a scenario and get an InjectObservation.
env = InjectArenaEnv(pg2=pg2, secalign=secalign, firewall=firewall, bank=bank)
# Cache attackers by source so we don't reload between calls.
attackers: Dict[str, Any] = {}
def get_attacker(source: str):
if source in attackers:
return attackers[source]
if source == "checkpoint":
attackers[source] = _load_attacker(args.checkpoint, args.max_new_tokens)
else:
attackers[source] = _load_zero_shot(args.max_new_tokens)
return attackers[source]
best: Optional[Dict[str, Any]] = None
best_score = -1
for attack_type, scenario_id in ATTACK_TYPE_TO_SCENARIO.items():
try:
scenario = bank.by_id(scenario_id)
except KeyError:
logger.warning("Scenario %s not in bank — skipping %s", scenario_id, attack_type)
continue
for steps in args.steps_labels:
source = "checkpoint" if steps > args.baseline_cutoff else "zero_shot"
model, tokenizer = get_attacker(source)
obs = env.reset(scenario_id=scenario_id)
payload = _generate_payload(model, tokenizer, obs, args.max_new_tokens, args.seed + steps)
pipe = _run_pipeline(payload, scenario, pg2, secalign, firewall)
trace = {
"attack_type": attack_type,
"steps": steps,
"scenario_id": scenario_id,
"scenario_label": scenario.get("target_behavior", ""),
"model_source": source,
"payload": payload,
**pipe,
}
out_path = out_dir / f"{attack_type}_{steps}.json"
with out_path.open("w") as f:
json.dump(trace, f, indent=2)
logger.info("Wrote %s (broke_pg2=%s, broke_fw=%s, task=%s)",
out_path.name,
pipe["outcome"]["broke_pg2"],
pipe["outcome"]["broke_fw"],
pipe["outcome"]["task_succeeded"])
# Score for highlight selection: composed > fw_bypass > pg2_bypass.
o = pipe["outcome"]
score = (4 if o["composed_bypass"] else 0) \
+ (2 if o["broke_fw"] else 0) \
+ (1 if o["broke_pg2"] else 0) \
+ (steps / 10000.0) # tiebreaker: prefer higher-step traces
if score > best_score:
best_score = score
best = trace
if best is not None:
with (highlight_dir / "highlight.json").open("w") as f:
json.dump(best, f, indent=2)
logger.info("Highlight: %s_%s (score=%.2f)", best["attack_type"], best["steps"], best_score)
logger.info("Done. Wrote %d traces to %s.", len(list(out_dir.glob("*.json"))), out_dir)
if __name__ == "__main__":
main()
|